use size_t type for proper large tensor support

src/layer/arm/convolution_3x3_winograd.h

@@ -4675,7 +4675,7 @@ static inline void conv3x3s1_winograd23_transform_input_tile(const Mat& bottom_b
     const int w = bottom_blob.w;
     const int h = bottom_blob.h;
     const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
+    const size_t N = bottom_blob.cstep * elempack;
 
     const int w_tiles = (w - 1) / 2;
 
@@ -5174,7 +5174,7 @@ static inline void conv3x3s1_winograd23_transform_output_tile(const Mat& top_til
     const int outw = top_blob.w;
     const int outh = top_blob.h;
     const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
+    const size_t N = top_blob.cstep * out_elempack;
 
     const int w_tiles = (outw + 1) / 2;
 
@@ -5819,7 +5819,7 @@ static inline void conv3x3s1_winograd43_transform_input_tile(const Mat& bottom_b
     const int w = bottom_blob.w;
     const int h = bottom_blob.h;
     const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
+    const size_t N = bottom_blob.cstep * elempack;
 
     const int w_tiles = (w + 1) / 4;
 
@@ -6620,7 +6620,7 @@ static inline void conv3x3s1_winograd43_transform_output_tile(const Mat& top_til
     const int outw = top_blob.w;
     const int outh = top_blob.h;
     const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
+    const size_t N = top_blob.cstep * out_elempack;
 
     const int w_tiles = (outw + 3) / 4;
 
@@ -7506,7 +7506,7 @@ static inline void conv3x3s1_winograd63_transform_input_tile(const Mat& bottom_b
     const int w = bottom_blob.w;
     const int h = bottom_blob.h;
     const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
+    const size_t N = bottom_blob.cstep * elempack;
 
     const int w_tiles = (w + 3) / 6;
 
@@ -8472,7 +8472,7 @@ static inline void conv3x3s1_winograd63_transform_output_tile(const Mat& top_til
     const int outw = top_blob.w;
     const int outh = top_blob.h;
     const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
+    const size_t N = top_blob.cstep * out_elempack;
 
     const int w_tiles = (outw + 5) / 6;
 

src/layer/arm/convolution_3x3_winograd_bf16s.h

@@ -13,7 +13,7 @@ static inline void conv3x3s1_winograd23_transform_input_tile_bf16s(const Mat& bo
     const int w = bottom_blob.w;
     const int h = bottom_blob.h;
     const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
+    const size_t N = bottom_blob.cstep * elempack;
 
     const int w_tiles = (w - 1) / 2;
 
@@ -514,7 +514,7 @@ static inline void conv3x3s1_winograd23_transform_output_tile_bf16s(const Mat& t
     const int outw = top_blob.w;
     const int outh = top_blob.h;
     const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
+    const size_t N = top_blob.cstep * out_elempack;
 
     const int w_tiles = (outw + 1) / 2;
 
@@ -1055,7 +1055,7 @@ static inline void conv3x3s1_winograd43_transform_input_tile_bf16s(const Mat& bo
     const int w = bottom_blob.w;
     const int h = bottom_blob.h;
     const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
+    const size_t N = bottom_blob.cstep * elempack;
 
     const int w_tiles = (w + 1) / 4;
 
@@ -1858,7 +1858,7 @@ static inline void conv3x3s1_winograd43_transform_output_tile_bf16s(const Mat& t
     const int outw = top_blob.w;
     const int outh = top_blob.h;
     const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
+    const size_t N = top_blob.cstep * out_elempack;
 
     const int w_tiles = (outw + 3) / 4;
 
@@ -2634,7 +2634,7 @@ static inline void conv3x3s1_winograd63_transform_input_tile_bf16s(const Mat& bo
     const int w = bottom_blob.w;
     const int h = bottom_blob.h;
     const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
+    const size_t N = bottom_blob.cstep * elempack;
 
     const int w_tiles = (w + 3) / 6;
 
@@ -3604,7 +3604,7 @@ static inline void conv3x3s1_winograd63_transform_output_tile_bf16s(const Mat& t
     const int outw = top_blob.w;
     const int outh = top_blob.h;
     const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
+    const size_t N = top_blob.cstep * out_elempack;
 
     const int w_tiles = (outw + 5) / 6;
 

src/layer/arm/convolution_3x3_winograd_int8.h

@@ -3626,7 +3626,7 @@ static inline void conv3x3s1_winograd23_transform_input_tile_int8(const Mat& bot
     const int w = bottom_blob.w;
     const int h = bottom_blob.h;
     const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
+    const size_t N = bottom_blob.cstep * elempack;
 
     const int w_tiles = (w - 1) / 2;
 
@@ -3919,7 +3919,7 @@ static inline void conv3x3s1_winograd23_transform_output_tile_int8(const Mat& to
     const int outw = top_blob.w;
     const int outh = top_blob.h;
     const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
+    const size_t N = top_blob.cstep * out_elempack;
 
     const int w_tiles = (outw + 1) / 2;
 
@@ -4484,7 +4484,7 @@ static inline void conv3x3s1_winograd43_transform_input_tile_int8(const Mat& bot
     const int w = bottom_blob.w;
     const int h = bottom_blob.h;
     const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
+    const size_t N = bottom_blob.cstep * elempack;
 
     const int w_tiles = (w + 1) / 4;
 
@@ -4896,7 +4896,7 @@ static inline void conv3x3s1_winograd43_transform_output_tile_int8(const Mat& to
     const int outw = top_blob.w;
     const int outh = top_blob.h;
     const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
+    const size_t N = top_blob.cstep * out_elempack;
 
     const int w_tiles = (outw + 3) / 4;
 

src/layer/arm/convolution_im2col_gemm.h

@@ -180,7 +180,7 @@ static void convolution_gemm_transB_packed_tile(const Mat& AT_tile, const Mat& B
     // NCNN_LOGE("convolution_gemm_transB_packed_tile %d %d %d %d %d %d", i, max_ii, j, max_jj, k, max_kk);
 
     const int out_elempack = top_blob.elempack;
-    const int out_hstep = (int)top_blob.cstep;
+    const size_t out_hstep = top_blob.cstep;
 
     const float* pAT = AT_tile;
     const float* pBT = BT_tile;

src/layer/arm/convolution_im2col_gemm_bf16s.h

@@ -6,7 +6,7 @@ static void convolution_gemm_transB_packed_tile_bf16s(const Mat& AT_tile, const
     // NCNN_LOGE("convolution_gemm_transB_packed_tile_bf16s %d %d %d %d %d %d", i, max_ii, j, max_jj, k, max_kk);
 
     const int out_elempack = top_blob.elempack;
-    const int out_hstep = (int)top_blob.cstep;
+    const size_t out_hstep = top_blob.cstep;
 
     const unsigned short* pAT = AT_tile;
     const unsigned short* pBT = BT_tile;

src/layer/arm/convolution_im2col_gemm_fp16s.h

@@ -6,7 +6,7 @@ static void convolution_gemm_transB_packed_tile_fp16sa(const Mat& AT_tile, const
     // NCNN_LOGE("convolution_gemm_transB_packed_tile_fp16sa %d %d %d %d %d %d", i, max_ii, j, max_jj, k, max_kk);
 
     const int out_elempack = top_blob.elempack;
-    const int out_hstep = (int)top_blob.cstep;
+    const size_t out_hstep = top_blob.cstep;
 
     const __fp16* pAT = AT_tile;
     const __fp16* pBT = BT_tile;

src/layer/arm/convolution_packed.h

@@ -499,14 +499,14 @@ static void convolution_packed(const Mat& bottom_blob, Mat& top_blob, const Mat&
     const int elempack = bottom_blob.elempack;
     const int inch = bottom_blob.c * elempack;
 
-    const int N = bottom_blob.cstep * elempack;
+    const size_t N = bottom_blob.cstep * elempack;
 
     const int outw = top_blob.w;
     const int outh = top_blob.h;
     const int out_elempack = top_blob.elempack;
     const int outch = top_blob.c * out_elempack;
 
-    const int M = top_blob.cstep * out_elempack;
+    const size_t M = top_blob.cstep * out_elempack;
 
     const int maxk = kernel_w * kernel_h;
 

src/layer/arm/convolution_packed_bf16s.h

@@ -499,14 +499,14 @@ static void convolution_packed_bf16s(const Mat& bottom_blob, Mat& top_blob, cons
     const int elempack = bottom_blob.elempack;
     const int inch = bottom_blob.c * elempack;
 
-    const int N = bottom_blob.cstep * elempack;
+    const size_t N = bottom_blob.cstep * elempack;
 
     const int outw = top_blob.w;
     const int outh = top_blob.h;
     const int out_elempack = top_blob.elempack;
     const int outch = top_blob.c * out_elempack;
 
-    const int M = top_blob.cstep * out_elempack;
+    const size_t M = top_blob.cstep * out_elempack;
 
     const int maxk = kernel_w * kernel_h;
 

src/layer/arm/convolution_packed_fp16s.h

@@ -448,14 +448,14 @@ static void convolution_packed_fp16s(const Mat& bottom_blob, Mat& top_blob, cons
     const int elempack = bottom_blob.elempack;
     const int inch = bottom_blob.c * elempack;
 
-    const int N = bottom_blob.cstep * elempack;
+    const size_t N = bottom_blob.cstep * elempack;
 
     const int outw = top_blob.w;
     const int outh = top_blob.h;
     const int out_elempack = top_blob.elempack;
     const int outch = top_blob.c * out_elempack;
 
-    const int M = top_blob.cstep * out_elempack;
+    const size_t M = top_blob.cstep * out_elempack;
 
     const int maxk = kernel_w * kernel_h;
 
@@ -1223,14 +1223,14 @@ static void convolution_packed_fp16sa(const Mat& bottom_blob, Mat& top_blob, con
     const int elempack = bottom_blob.elempack;
     const int inch = bottom_blob.c * elempack;
 
-    const int N = bottom_blob.cstep * elempack;
+    const size_t N = bottom_blob.cstep * elempack;
 
     const int outw = top_blob.w;
     const int outh = top_blob.h;
     const int out_elempack = top_blob.elempack;
     const int outch = top_blob.c * out_elempack;
 
-    const int M = top_blob.cstep * out_elempack;
+    const size_t M = top_blob.cstep * out_elempack;
 
     const int maxk = kernel_w * kernel_h;
 

src/layer/arm/convolution_packed_int8.h

@@ -536,7 +536,7 @@ static void convolution_packed_int8(const Mat& bottom_blob, Mat& top_blob, const
     const int elempack = bottom_blob.elempack;
     const int inch = bottom_blob.c * elempack;
 
-    const int N = bottom_blob.cstep * elempack;
+    const size_t N = bottom_blob.cstep * elempack;
 
     const int outw = top_blob.w;
     const int outh = top_blob.h;
@@ -576,8 +576,8 @@ static void convolution_packed_int8(const Mat& bottom_blob, Mat& top_blob, const
         // shadowed variable for less openmp task args
         const int outw = top_blob.w;
         const int outh = top_blob.h;
-        const int N = bottom_blob.cstep * elempack;
-        const int M = top_blob.cstep * out_elempack;
+        const size_t N = bottom_blob.cstep * elempack;
+        const size_t M = top_blob.cstep * out_elempack;
 
         int* outptr = top_blob.channel(p / out_elempack);
 
@@ -908,8 +908,8 @@ static void convolution_packed_int8(const Mat& bottom_blob, Mat& top_blob, const
         // shadowed variable for less openmp task args
         const int outw = top_blob.w;
         const int outh = top_blob.h;
-        const int N = bottom_blob.cstep * elempack;
-        const int M = top_blob.cstep * out_elempack;
+        const size_t N = bottom_blob.cstep * elempack;
+        const size_t M = top_blob.cstep * out_elempack;
 
         int* outptr = top_blob.channel(p / out_elempack);
 
@@ -1167,7 +1167,7 @@ static void convolution_packed_int8(const Mat& bottom_blob, Mat& top_blob, const
         // shadowed variable for less openmp task args
         const int outw = top_blob.w;
         const int outh = top_blob.h;
-        const int N = bottom_blob.cstep * elempack;
+        const size_t N = bottom_blob.cstep * elempack;
 
         int* outptr0 = top_blob.channel(p);
         int* outptr1 = top_blob.channel(p + 1);

src/layer/arm/gemm_arm.cpp

@@ -45,7 +45,7 @@ Gemm_arm::Gemm_arm()
 void pack_A_tile(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
 {
     const int elempack = A.elempack;
-    const int A_hstep = A.dims == 3 ? (int)A.cstep : A.w;
+    const size_t A_hstep = A.dims == 3 ? A.cstep : (size_t)A.w;
 
     float* pp = AT;
 
@@ -256,7 +256,7 @@ void pack_A_tile(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
 static void transpose_pack_A_tile(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
 {
     const int elempack = A.elempack;
-    const int A_hstep = A.dims == 3 ? (int)A.cstep : A.w;
+    const size_t A_hstep = A.dims == 3 ? A.cstep : (size_t)A.w;
 
     float* pp = AT;
 
@@ -400,7 +400,7 @@ static void transpose_pack_A_tile(const Mat& A, Mat& AT, int i, int max_ii, int
 static void pack_B_tile(const Mat& B, Mat& BT, int j, int max_jj, int k, int max_kk)
 {
     const int elempack = B.elempack;
-    const int B_hstep = B.dims == 3 ? (int)B.cstep : B.w;
+    const size_t B_hstep = B.dims == 3 ? B.cstep : (size_t)B.w;
 
     float* pp = BT;
 
@@ -708,7 +708,7 @@ static void pack_B_tile(const Mat& B, Mat& BT, int j, int max_jj, int k, int max
 static void transpose_pack_B_tile(const Mat& B, Mat& BT, int j, int max_jj, int k, int max_kk)
 {
     const int elempack = B.elempack;
-    const int B_hstep = B.dims == 3 ? (int)B.cstep : B.w;
+    const size_t B_hstep = B.dims == 3 ? B.cstep : (size_t)B.w;
 
     float* pp = BT;
 
@@ -895,7 +895,7 @@ static void transpose_pack_B_tile(const Mat& B, Mat& BT, int j, int max_jj, int
 static void transpose_unpack_output_tile(const Mat& topT, Mat& top_blob, int i, int max_ii, int j, int max_jj)
 {
     const int out_elempack = top_blob.elempack;
-    const int out_hstep = top_blob.dims == 3 ? (int)top_blob.cstep : top_blob.w;
+    const size_t out_hstep = top_blob.dims == 3 ? top_blob.cstep : (size_t)top_blob.w;
 
     const float* pp = topT;
 
@@ -1042,7 +1042,7 @@ static void transpose_unpack_output_tile(const Mat& topT, Mat& top_blob, int i,
 static void gemm_transB_packed_tile(const Mat& AT_tile, const Mat& BT_tile, const Mat& CT_tile, Mat& topT_tile, Mat& top_blob, int broadcast_type_C, int i, int max_ii, int j, int max_jj, int k, int max_kk, bool k_end)
 {
     const int out_elempack = top_blob.elempack;
-    const int out_hstep = top_blob.dims == 3 ? (int)top_blob.cstep : top_blob.w;
+    const size_t out_hstep = top_blob.dims == 3 ? top_blob.cstep : (size_t)top_blob.w;
 
     const float* pAT = AT_tile;
     const float* pBT = BT_tile;

src/layer/arm/gemm_arm_asimdhp.cpp

@@ -23,7 +23,7 @@ namespace ncnn {
 static void gemm_transB_packed_tile_fp16sa(const Mat& AT_tile, const Mat& BT_tile, const Mat& CT_tile, Mat& topT_tile, Mat& top_blob, int broadcast_type_C, int i, int max_ii, int j, int max_jj, int k, int max_kk, bool k_end)
 {
     const int out_elempack = top_blob.elempack;
-    const int out_hstep = top_blob.dims == 3 ? (int)top_blob.cstep : top_blob.w;
+    const size_t out_hstep = top_blob.dims == 3 ? top_blob.cstep : (size_t)top_blob.w;
 
     const __fp16* pAT = AT_tile;
     const __fp16* pBT = BT_tile;

src/layer/arm/gemm_bf16s.h

@@ -3,7 +3,7 @@
 
 static void pack_A_tile_fp32_to_bf16(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
 {
-    const int A_hstep = A.dims == 3 ? (int)A.cstep : A.w;
+    const size_t A_hstep = A.dims == 3 ? A.cstep : (size_t)A.w;
 
     unsigned short* pp = AT;
 
@@ -192,7 +192,7 @@ static void pack_A_tile_fp32_to_bf16(const Mat& A, Mat& AT, int i, int max_ii, i
 
 static void transpose_pack_A_tile_fp32_to_bf16(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
 {
-    const int A_hstep = A.dims == 3 ? (int)A.cstep : A.w;
+    const size_t A_hstep = A.dims == 3 ? A.cstep : (size_t)A.w;
 
     unsigned short* pp = AT;
 
@@ -256,7 +256,7 @@ static void transpose_pack_A_tile_fp32_to_bf16(const Mat& A, Mat& AT, int i, int
 
 static void pack_B_tile_fp32_to_bf16(const Mat& B, Mat& BT, int j, int max_jj, int k, int max_kk)
 {
-    const int B_hstep = B.dims == 3 ? (int)B.cstep : B.w;
+    const size_t B_hstep = B.dims == 3 ? B.cstep : (size_t)B.w;
 
     unsigned short* pp = BT;
 
@@ -567,7 +567,7 @@ static void pack_B_tile_fp32_to_bf16(const Mat& B, Mat& BT, int j, int max_jj, i
 
 static void transpose_pack_B_tile_fp32_to_bf16(const Mat& B, Mat& BT, int j, int max_jj, int k, int max_kk)
 {
-    const int B_hstep = B.dims == 3 ? (int)B.cstep : B.w;
+    const size_t B_hstep = B.dims == 3 ? B.cstep : (size_t)B.w;
 
     unsigned short* pp = BT;
 
@@ -646,7 +646,7 @@ static void transpose_pack_B_tile_fp32_to_bf16(const Mat& B, Mat& BT, int j, int
 static void transpose_unpack_output_tile_fp32_to_bf16(const Mat& topT, Mat& top_blob, int i, int max_ii, int j, int max_jj)
 {
     const int out_elempack = top_blob.elempack;
-    const int out_hstep = top_blob.dims == 3 ? (int)top_blob.cstep : top_blob.w;
+    const size_t out_hstep = top_blob.dims == 3 ? top_blob.cstep : (size_t)top_blob.w;
 
     const float* pp = topT;
 
@@ -785,7 +785,7 @@ static void transpose_unpack_output_tile_fp32_to_bf16(const Mat& topT, Mat& top_
 static void gemm_transB_packed_tile_bf16s(const Mat& AT_tile, const Mat& BT_tile, const Mat& CT_tile, Mat& topT_tile, Mat& top_blob, int broadcast_type_C, float alpha, int i, int max_ii, int j, int max_jj, int k, int max_kk, bool k_end)
 {
     const int out_elempack = top_blob.elempack;
-    const int out_hstep = top_blob.dims == 3 ? (int)top_blob.cstep : top_blob.w;
+    const size_t out_hstep = top_blob.dims == 3 ? top_blob.cstep : (size_t)top_blob.w;
 
     const unsigned short* pAT = AT_tile;
     const unsigned short* pBT = BT_tile;